def extract(node):
# update the attributes of the node
block_size = node.pb.attr['block_size'].i
data_format = node.pb.attr['data_format'].s.decode('utf-8')
Op.get_op_class_by_name(__class__.op).update_node_stat(node,
{'block_size': block_size, 'data_format': data_format})
return __class__.enabled
def extract(node):
param = node.pb.python_param
attrs = CaffePythonFrontExtractorOp.parse_param_str(param.param_str)
update_attrs = {
'feat_stride': 16,
'base_size': 16,
'min_size': 16,
'ratio': [0.5, 1, 2],
'scale': [8, 16, 32],
'pre_nms_topn': 6000,
'post_nms_topn': 300,
'nms_thresh': 0.7
}
if 'ratios' in attrs and 'ratio' in attrs:
log.error(
'Both ratios and ratio found, value of ratios will be used',
extra={'is_warning': True})
if 'scales' in attrs and 'scale' in attrs:
log.error(
'Both scales and scale found, value of scales will be used',
extra={'is_warning': True})
if 'ratios' in attrs:
attrs['ratio'] = attrs['ratios']
del attrs['ratios']
if 'scales' in attrs:
attrs['scale'] = attrs['scales']
del attrs['scales']
update_attrs.update(attrs)
CaffePythonFrontExtractorOp.check_param(
Op.get_op_class_by_name('Proposal'), update_attrs)
Op.get_op_class_by_name('Proposal').update_node_stat(
node, update_attrs)
return __class__.enabled
def extract(node):
proto_layer = node.pb
param = proto_layer.region_yolo_param
flatten_param = proto_layer.flatten_param
axis = flatten_param.axis
end_axis = flatten_param.end_axis
coords = param.coords
classes = param.classes
num = param.num
update_attrs = {
'coords': coords,
'classes': classes,
'num': num,
'do_softmax': int(param.do_softmax),
'anchors': np.array(param.anchors),
'mask': np.array(param.mask)
}
flatten_attrs = {'axis': axis, 'end_axis': end_axis}
mapping_rule = merge_attrs(param, update_attrs)
mapping_rule.update(flatten_attrs)
mapping_rule.update(layout_attrs())
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(
node, mapping_rule)
return __class__.enabled
def extract(node):
proto_layer = node.pb
param = proto_layer.prior_box_param
variance = param.variance
if len(variance) == 0:
variance = [0.1]
update_attrs = {
'aspect_ratio': np.array(param.aspect_ratio),
'min_size': np.array(param.min_size),
'max_size': np.array(param.max_size),
'flip': int(param.flip),
'clip': int(param.clip),
'variance': list(variance),
'img_size': param.img_size,
'img_h': param.img_h,
'img_w': param.img_w,
'step': param.step,
'step_h': param.step_h,
'step_w': param.step_w,
'offset': param.offset,
}
mapping_rule = merge_attrs(param, update_attrs)
mapping_rule.update(layout_attrs())
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(
node, mapping_rule)
return __class__.enabled
def extract(node):
proto_layer = node.pb
param = proto_layer.augmentation_param
# slice_dim is deprecated parameter and is used as alias for axis
# however if slice_dim is defined and axis is default, we use slice_dim
update_attrs = {
'crop_width': param.crop_width,
'crop_height': param.crop_height,
'write_augmented': param.write_augmented,
'max_multiplier': param.max_multiplier,
'augment_during_test': int(param.augment_during_test),
'recompute_mean': param.recompute_mean,
'write_mean': param.write_mean,
'mean_per_pixel': int(param.mean_per_pixel),
'mean': param.mean,
'mode': param.mode,
'bottomwidth': param.bottomwidth,
'bottomheight': param.bottomheight,
'num': param.num,
'chromatic_eigvec': param.chromatic_eigvec
}
mapping_rule = merge_attrs(param, update_attrs)
if node.model_pb:
for index in range(0, len(node.model_pb.blobs)):
embed_input(mapping_rule, index + 1, 'custom_{}'.format(index),
node.model_pb.blobs[index].data)
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(
node, mapping_rule)
return __class__.enabled
def check_init_states(graph: Graph, match: dict):
"""
Check if cell have initial states and create zeros states if not.
"""
rnn_layer = match['rnn_layer']
num_directions = 2 if rnn_layer.direction == 'bidirectional' else 1
batch_size = rnn_layer.in_node(0).shape[rnn_layer.batch_dim]
h_init_port = 5
c_init_port = 6
if h_init_port not in rnn_layer.in_nodes():
h_shape = [num_directions, batch_size, rnn_layer.hidden_size] # from ONNX spec
h_init = np.full(h_shape, 0, dtype=np.float32)
Op.create_and_connect_input_data_node(
graph,
rnn_layer,
{'value': h_init, 'shape': np.array(h_init.shape, dtype=np.int64)},
{'in': h_init_port, 'permutation': None}
)
if rnn_layer.op == 'LSTM':
if c_init_port not in rnn_layer.in_nodes():
c_shape = [num_directions, batch_size, rnn_layer.hidden_size] # from ONNX spec
c_init = np.full(c_shape, 0, dtype=np.float32)
Op.create_and_connect_input_data_node(
graph,
rnn_layer,
{'value': c_init, 'shape': np.array(c_init.shape, dtype=np.int64)},
{'in': c_init_port, 'permutation': None}
)
def apply_scale(graph: Graph, input_node: Node,
node_mean_scale_values: dict):
if 'scale' in node_mean_scale_values and node_mean_scale_values[
'scale'] is not None:
if all([x == 1 for x in node_mean_scale_values['scale']]):
return
out_node = input_node.out_node()
if not input_node.has_valid('shape'):
raise Error("Node {} has not valid shape attribute".format(
input_node.id))
input_shape = input_node.shape
# Create Mul node
value = 1 / np.array(node_mean_scale_values['scale'])
graph.remove_edge(input_node.id, out_node.id)
mul_node = Mul(graph, dict(name="Mul_"))
mul_data = Op.create_input_data_node(graph, "data_mul_",
np.array(value))
Op.expand_node_shape(mul_data,
(len(input_shape) -
2 if graph.graph['layout'] == 'NCHW' else 0))
mul_input = Op.create_data_node(graph, input_node,
{'shape': out_node.shape})
mul_node.create_node_with_data(inputs=[mul_input, mul_data],
data_nodes=out_node)
def extract_proposal_params(node, defaults):
param = node.pb.python_param
attrs = CaffePythonFrontExtractorOp.parse_param_str(param.param_str)
update_attrs = defaults
if 'ratios' in attrs and 'ratio' in attrs:
log.error(
'Both ratios and ratio found, value of ratios will be used',
extra={'is_warning': True})
if 'scales' in attrs and 'scale' in attrs:
log.error(
'Both scales and scale found, value of scales will be used',
extra={'is_warning': True})
if 'ratios' in attrs:
attrs['ratio'] = attrs['ratios']
del attrs['ratios']
if 'scales' in attrs:
attrs['scale'] = attrs['scales']
del attrs['scales']
update_attrs.update(attrs)
CaffePythonFrontExtractorOp.check_param(
Op.get_op_class_by_name('Proposal'), update_attrs)
Op.get_op_class_by_name('Proposal').update_node_stat(
node, update_attrs)
def _scale_input_action_mul(graph: nx.MultiDiGraph, match: dict, scale: float):
assert (len(match['placeholder'].out_nodes()))
tinput = match['placeholder']
if not tinput.has_valid('shape'):
raise Error("Node {} has not valid shape attribute".format(tinput.id))
input_shape = tinput.shape
toutput = match['data']
# Create Mul node
value = np.array([1 / scale])
# Disconnect input with data node
graph.remove_edge(tinput.id, toutput.id)
# Create Mul node
mul_node = Mul(graph, dict(name="Mul1_"))
mul_data = Op.create_input_data_node(graph, "data_mul_scale_",
np.array(value))
Op.expand_node_shape(
mul_data,
len(input_shape) - 2 if graph.graph['layout'] == 'NCHW' else 0)
mul_input = Op.create_data_node(graph, tinput, {'shape': toutput.shape})
mul_node.create_node_with_data(inputs=[mul_input, mul_data],
data_nodes=toutput)
def extract(node):
proto_layer = node.pb
param = proto_layer.resample_param
types = [
"", 'caffe.ResampleParameter.NEAREST',
'caffe.ResampleParameter.LINEAR', 'caffe.ResampleParameter.CUBIC',
'caffe.ResampleParameter.AREA'
]
resample_type = types[param.type]
update_attrs = {
'antialias': int(param.antialias),
'height': param.height,
'width': param.width,
'type': resample_type,
'factor': param.factor
}
mapping_rule = merge_attrs(param, update_attrs)
mapping_rule['resample_type'] = mapping_rule['type']
mapping_rule.pop('type')
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(
node, mapping_rule)
return __class__.enabled
def apply_mean_value(graph: Graph, input_node: Node,
node_mean_scale_values: dict):
if 'mean' in node_mean_scale_values and node_mean_scale_values[
'mean'] is not None:
if all([x == 0 for x in node_mean_scale_values['mean']]):
return
out_node = input_node.out_node()
if not input_node.has_valid('shape'):
raise Error("Node {} has not valid shape attribute".format(
input_node.id))
input_shape = input_node.shape
# Create Add node
graph.remove_edge(input_node.id, out_node.id)
value = np.array(node_mean_scale_values['mean']) * (-1)
add_node = Add(graph, dict(name="Add_"))
add_data = Op.create_input_data_node(graph, "data_add_",
np.array(value))
Op.expand_node_shape(add_data,
(len(input_shape) -
2 if graph.graph['layout'] == 'NCHW' else 0))
add_input = Op.create_data_node(graph, input_node,
{'shape': out_node.shape})
add_node.create_node_with_data(inputs=[add_input, add_data],
data_nodes=out_node)
def split_bilstm(self, bilstm, new_init_hiddens, new_init_cells,
splitted_W, splitted_R, splitted_B):
""" Split one bilstm node into 2 one-directional lstm nodes.
All input data nodes should be already prepared; they are
have 2 in the major dimension.
"""
assert len(bilstm.out_nodes()) == 3
all_outputs = []
for i in [0, 1]:
direction = ['forward', 'reverse'][i]
op = LSTMSequence(
bilstm.graph, {
'hidden_size': bilstm.hidden_size,
'direction': direction,
'batch_dim': bilstm.batch_dim,
'sequence_dim': bilstm.sequence_dim,
'blobs_wrb': bilstm.blobs_wrb,
'has_num_directions': bilstm.has_num_directions,
'format': bilstm.format,
'name': bilstm.name + '/Split/' + direction,
})
output_data = Op._create_data_node(
bilstm.graph,
name=bilstm.out_node(0).name + '/Split/' + str(i),
attrs={'shape': bilstm.out_node(0).shape.copy()})
assert output_data.shape[1] == 2
output_data.shape[1] = 1
output_hidden = Op._create_data_node(
bilstm.graph,
name=bilstm.out_node(1).name + '/Split/' + str(i),
attrs={'shape': bilstm.out_node(1).shape.copy()})
assert output_hidden.shape[0] == 2
output_hidden.shape[0] = 1
output_cell = Op._create_data_node(
bilstm.graph,
name=bilstm.out_node(2).name + '/Split/' + str(i),
attrs={'shape': bilstm.out_node(2).shape.copy()})
assert output_cell.shape[0] == 2
output_cell.shape[0] = 1
all_outputs.append(
op.create_node_with_data(
inputs=[
bilstm.in_node(0),
splitted_W[i],
splitted_R[i],
splitted_B[i],
None,
new_init_hiddens[i],
new_init_cells[i],
],
data_nodes=[output_data, output_hidden, output_cell]))
return all_outputs
def extract(node):
attrs = get_mxnet_layer_attrs(node.symbol_dict)
pre_nms_topn = attrs.int('rpn_pre_nms_top_n', 6000)
post_nms_topn = attrs.int('rpn_post_nms_top_n', 300)
nms_thresh = attrs.float('threshold', 0.7)
min_size = attrs.int('rpn_min_size', 16)
scale = attrs.tuple("scales", float, (4, 8, 16, 32))
ratio = attrs.tuple("ratios", float, (0.5, 1, 2))
feat_stride = attrs.int('feature_stride', 16)
update_attrs = {
'feat_stride': feat_stride,
'ratio': np.array(ratio),
'min_size': min_size,
'scale': np.array(scale),
'pre_nms_topn': pre_nms_topn,
'post_nms_topn': post_nms_topn,
'nms_thresh': nms_thresh,
'base_size': feat_stride
}
# update the attributes of the node
Op.get_op_class_by_name('Proposal').update_node_stat(
node, update_attrs)
return __class__.enabled
def extract(node):
proto_layer = node.pb
pb_model = node.model_pb
param = proto_layer.prelu_param
update_attrs = {
'channel_shared': int(param.channel_shared)
}
variance_norm_caffe_map = {
0: 'caffe.FillerParameter.FAN_IN',
1: 'caffe.FillerParameter.FAN_OUT',
2: 'caffe.FillerParameter.AVERAGE'
}
if hasattr(param, 'filler'):
update_attrs.update({
'filler_type': param.filler.type,
'filler_value': int(param.filler.value),
'min': int(param.filler.min),
'max': int(param.filler.max),
'mean': int(param.filler.mean),
'std': int(param.filler.std),
'sparse': param.filler.sparse,
'variance_norm': variance_norm_caffe_map[param.filler.variance_norm]
})
mapping_rule = merge_attrs(param, update_attrs)
mapping_rule.update(weights_biases(False, pb_model))
mapping_rule.update(layout_attrs())
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(node, mapping_rule)
return __class__.enabled
def _fused_batch_norm_decomposition(graph: Graph, tinput: Node, toutput: Node, gamma: Node, beta: Node,
mean: np.ndarray, variance: np.ndarray, can_be_fused=True):
"""
This is common function for TF, Caffe and MXNet
It creates Mul->Add->Mul->Add subgraph
"""
shape = tinput.shape
# Create first Mul & Add operations
mul1_node = Mul(graph, dict(name="Mul1_", can_be_fused=can_be_fused))
add1_node = Add(graph, dict(name="Add1_", can_be_fused=can_be_fused))
mul1_data = Op.create_input_data_node(graph, "data_mul_", np.array(mean))
add1_data = Op.create_input_data_node(graph, "data_add_", np.array(variance))
# Broadcast const from scalar
# We can broadcast only when const.value is scalar
if gamma.shape[0] != gamma.value.shape[0]:
gamma.value.resize(gamma.shape)
gamma.value.fill(gamma.value[0])
# Create second Mul & Add
mul2_node = Mul(graph, dict(name="Mul2_", can_be_fused=can_be_fused))
add2_node = Add(graph, dict(name="Add2_", can_be_fused=can_be_fused))
add2_node.create_node_with_data(
inputs=[mul2_node.create_node_with_data(
inputs=[add1_node.create_node_with_data(
inputs=[mul1_node.create_node_with_data(inputs=[tinput, mul1_data]),
add1_data]),
gamma]),
beta],
data_nodes=toutput)
def extract(node):
proto_layer = node.pb
param = proto_layer.correlation_param
corr_type = 'caffe.CorrelationParameter.MULTIPLY'
if param.correlation_type == 1:
corr_type = 'caffe.CorrelationParameter.SUBTRACT'
update_attrs = {
'pad': param.pad,
'kernel_size': param.kernel_size,
'max_displacement': param.max_displacement,
'stride_1': param.stride_1,
'stride_2': param.stride_2,
'single_direction': param.single_direction,
'do_abs': int(param.do_abs),
'correlation_type': corr_type,
}
mapping_rule = merge_attrs(param, update_attrs)
mapping_rule.update(layout_attrs())
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(
node, mapping_rule)
return __class__.enabled
def extract(node):
proto_layer = node.pb
param = proto_layer.accum_param
attrs = collect_attributes(param)
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(node, attrs)
return __class__.enabled
def extract(node):
mapping_rule = collect_attributes(node.pb.shuffle_channel_param)
mapping_rule.update(layout_attrs())
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(
node, mapping_rule)
return __class__.enabled
def extract(node):
attrs = {
'data_type': tf_dtype_extractor(node.pb.attr["dtype"].type),
'shape': tf_tensor_shape(node.pb.attr["shape"].shape),
'identity': True,
}
Op.update_node_stat(node, attrs)
return __class__.enabled
def convert_batch_norm(graph: nx.MultiDiGraph):
"""
This function finds FusedBatchNorm layer (or BatchNorm for MXNet) and replaces with Mul->Add->Mul->Add sequence.
"""
for n in list(graph.nodes()):
node = Node(graph, n)
if node.has_valid('op') and (node.op == 'FusedBatchNorm'
or node.op == 'BatchNorm'
or node.op == 'BatchNormalization'):
toutput = node.out_node()
tinput = node.in_node(0)
if any([
node.in_node(i).value is None
for i in range(1, len(node.in_nodes()))
]):
log.warning(
'Cannot translate FusedBatchNorm {} node with non-constant weights'
.format(
node.name if node.has_valid('name') else '<UNKNOWN>'))
continue
const = node.in_node(1)
beta = node.in_node(2)
mean = node.in_node(3)
variance = node.in_node(4)
eps = node.eps
if node.has_valid('fix_gamma') and node.fix_gamma:
const.value.fill(1.)
can_be_fused = False if not node.soft_get('can_be_fused') else True
# Remove edges from FusedBN node
graph.remove_edge(tinput.id, node.id)
graph.remove_edge(beta.id, node.id)
graph.remove_edge(const.id, node.id)
graph.remove_edge(mean.id, node.id)
graph.remove_edge(variance.id, node.id)
graph.remove_edge(node.id, toutput.id)
scale = 1. / np.sqrt(variance.value + eps)
shift = (mean.value * (-1)) * scale
# Expand dims for current layout
broadcast_dims_cnt = len(
tinput.shape) - 2 if graph.graph['layout'] == 'NCHW' else 0
# Update values and shapes with new shape
Op.expand_node_shape(const, broadcast_dims_cnt)
Op.expand_node_shape(beta, broadcast_dims_cnt)
for idx in range(broadcast_dims_cnt):
scale = np.expand_dims(scale, axis=-1)
shift = np.expand_dims(shift, axis=-1)
_fused_batch_norm_decomposition(graph, tinput, toutput, const,
beta, scale, shift, can_be_fused)
def repack_weights(self, graph: Graph, match: dict):
# Concat W, R in IE- format
# Delete useless num_dir dimensions and n_cells dimensions in W, R, B (peepholes?)
lstm = match['rnn_layer']
W, R, B = match['W'].value.copy(), match['R'].value.copy(), match['B'].value.copy()
graph.remove_edge(match['W'].id, lstm.id)
graph.remove_edge(match['R'].id, lstm.id)
graph.remove_edge(match['B'].id, lstm.id)
# Sum component of B that correspond to W and R
if lstm.op == 'GRU' and lstm.linear_before_reset:
B_shape = np.array(B.shape)
B_shape[3] = 4
B_shape[2] = 1
B_tmp = np.zeros(shape=B_shape)
B_tmp[:, :, :, 0, :] = B[:, :, 0, 0, :] + B[:, :, 1, 0, :]
B_tmp[:, :, :, 1, :] = B[:, :, 0, 1, :] + B[:, :, 1, 1, :]
B_tmp[:, :, :, 2, :] = B[:, :, 0, 2, :][:, :, np.newaxis, :]
B_tmp[:, :, :, 3, :] = B[:, :, 1, 2, :][:, :, np.newaxis, :]
B = B_tmp
else:
B = np.add.reduce(B, axis=2, keepdims=True)
# Concatenate W, R to IE-compatible format
assert len(W.shape) == 5
assert len(R.shape) == 5
WR = np.concatenate([W, R], axis=4)
# Squeeze useless dimensions
assert WR.shape[0] == 1 # num_dir == 1
assert WR.shape[1] == 1 # num_cells == 1
assert B.shape[0] == 1
assert B.shape[1] == 1
WR = WR.squeeze(axis=(0, 1))
B = B.squeeze(axis=(0, 1))
# Flatten all output (0, 1) and input dimensions (2, 3)
final_shape_WR = [WR.shape[0] * WR.shape[1], -1]
assert final_shape_WR[0] == lstm.hidden_size * lstm.multiplier
WR = WR.reshape(final_shape_WR)
final_shape_B = final_shape_WR
if lstm.op == 'GRU' and lstm.linear_before_reset:
final_shape_B[0] = lstm.hidden_size * 4
B = B.reshape(final_shape_B)
# Squeeze fake dimension in B
B = B.squeeze(axis=-1)
for blob, port, name in [(WR, 1, 'weights'), (B, 2, 'biases')]:
Op.create_and_connect_input_data_node(
graph,
lstm,
{'value': blob, 'shape': np.array(blob.shape, dtype=np.int64)},
{'in': port, 'bin': name, 'permutation': None}
)
def extract(node):
proto_layer = node.pb
param = proto_layer.norm_param
attrs = collect_attributes(param, enable_flattening_nested_params=True)
attrs.update(weights_biases(False, node.model_pb))
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(node, attrs)
return __class__.enabled
def extract(node):
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(
node, {
'out_max_val': 0,
'top_k': 1,
'axis': None,
'dim_attrs': ['axis']
})
return __class__.enabled
def extract(cls, node):
attrs = {
'data_type': tf_dtype_extractor(node.pb.attr["dtype"].type),
'shape': tf_tensor_shape(node.pb.attr["shape"].shape),
'identity': True,
'infer':
lambda node: copy_shape_infer(node, value_infer=copy_value),
}
Op.update_node_stat(node, attrs)
return cls.enabled
def extract(node):
proto_layer = node.pb
param = proto_layer.attr
# extracting parameters from TensorFlow layer and prepare them for IR
attrs = {'op': __class__.op}
# update the attributes of the node
Op.get_op_class_by_name(__class__.op).update_node_stat(node, attrs)
return __class__.enabled
def extract(cls, node):
shapes = node.pb.attr['output_shapes'].list.shape
tf_types = node.pb.attr['output_types'].list.type
extracted_types = []
for t in tf_types:
extracted_types.append(tf_dtype_extractor(t))
result_shapes = []
for shape_pb in shapes:
result_shapes.append(tf_tensor_shape(shape_pb))
Op.update_node_stat(node, {'shapes': result_shapes, 'types': extracted_types, 'out_ports_count': 1})
return cls.enabled
def extract(cls, node):
narrow_range = node.pb.attr['narrow_range'].b
num_bits = node.pb.attr['num_bits'].i
levels = 2 ** num_bits - int(narrow_range)
# we prepare this operation to be converted to FakeQuantize op,
# but input reconnection is needed, so we don't set infer function and type attribute
Op.update_node_stat(node, {'op': 'FakeQuantWithMinMaxVars', 'levels': levels,
'narrow_range': narrow_range, 'num_bits': num_bits})
return cls.enabled
def _create_data_if_necessary(self):
if self.node.graph.stage == 'front':
raise Error("_create_data_if_necessary method is not applicable for front Graph phase!")
if self.type == 'in':
raise Error("_create_data_if_necessary method is not applicable for 'in' Port type!")
if self.idx not in self.node.out_nodes(control_flow=self.control_flow):
from mo.ops.op import Op
Op.create_data_node(self.node.graph, self.node, out_port=self.idx)
self.node['need_shape_inference'] = True
return self.node.out_node(self.idx, control_flow=self.control_flow)
def copy_input_blobs(op: Node, copy_op: Node):
"""
Function copy input blob data nodes from restored graph to copied one
:param op: Node from restored graph
:param copy_op: Node from copied graph
:return:
"""
for u, d in op.get_sorted_inputs():
if 'bin' in d:
Op.create_and_connect_input_data_node(copy_op.graph, copy_op,
{'value': op.in_node(d['in']).value,
'shape': op.in_node(d['in']).shape}, d)
def repack_weights(graph: Graph, match: dict):
"""
Repack weights into general format (described above) and reorder gates.
"""
rnn_layer = match['rnn_layer']
W = match['W'].value.copy()
R = match['R'].value.copy()
num_directions = 2 if rnn_layer.direction == 'bidirectional' else 1
graph.remove_edge(match['W'].id, rnn_layer.id)
graph.remove_edge(match['R'].id, rnn_layer.id)
# find optional 'B' biases blob
if 3 in rnn_layer.in_nodes():
# TODO: check if 'bin': 'B' attribute is assigned to this edge
B = rnn_layer.in_node(3).value.copy()
graph.remove_edge(rnn_layer.in_node(3).id, rnn_layer.id)
else:
B_shape = [num_directions, 2 * rnn_layer.multiplier * rnn_layer.hidden_size] # from ONNX spec
B = np.full(B_shape, 0, dtype=np.float32)
# Add extra dimensions for W, R and B for easier repacking and reordering
B = B.reshape([
num_directions, # 0: num of directions
rnn_layer.num_layers, # 1: num_layers
2, # 2: two input parts of the matrix: W, R
rnn_layer.multiplier, # 3: four output parts of the matrix for all gates in order: i, o, f, c
rnn_layer.hidden_size, # 4: output size per direction and gate
])
W, R = [x.reshape([
num_directions, # 0: num of directions
rnn_layer.num_layers, # 1: num_layers
rnn_layer.multiplier, # 2: four output parts of the matrix for all gates in order: i, o, f, c
rnn_layer.hidden_size, # 3: output size per direction and gate
-1]) # 4: input size/hidden size in W/R correspondingly
for x in (W, R)]
input_size = match['input'].shape[2]
assert input_size == W.shape[-1]
# Reorder gates: iofc --> fico
gate_reorder = rnn_layer.gate_order
W, R = (np.take(x, gate_reorder, axis=2) for x in (W, R))
B = np.take(B, gate_reorder, axis=3)
for blob, port in [(W, 1), (R, 2), (B, 3)]:
Op.create_and_connect_input_data_node(
graph,
rnn_layer,
{'value': blob, 'shape': np.array(blob.shape, dtype=np.int64)},
{'in': port, 'permutation': None}
)
